Automatic Cloud Segmentation and Cloud Cover Retrieval from Wide-Field Thermal Infrared Whole Sky Images

Thermal infrared (TIR) whole-sky imaging enables continuous cloud observations during both daytime and nighttime, making it an important tool for ground-based cloud monitoring. However, existing infrared cloud detection methods mainly rely on manually selected thresholds or handcrafted image features, which typically require instrument-specific calibration and exhibit limited generalization capability. Moreover, wide-field imaging covers not only the zenith region but also the peripheral areas at large zenith angles, which are susceptible to the combined effects of strong atmospheric background emission, thermal radiation contamination from the surface and surrounding environment, and imaging geometric distortion, resulting in reduced cloud–background contrast and making cloud segmentation considerably more challenging than in conventional infrared sky imagery. To address these limitations, this study proposes a deep-learning framework for automatic cloud detection and cloud cover retrieval from wide-field TIR whole-sky images. The framework adopts a two-stage strategy in which infrared images are first classified as clear-sky or cloudy, followed by semantic segmentation of cloudy images for cloud cover estimation. To support model training and evaluation, a dataset containing 3000 wide-field TIR whole-sky images was established using a semi-automatic labeling procedure. Experimental results demonstrate that the proposed method effectively identifies cloud structures, including those located near the edge of wide-field whole-sky images. The classifier achieved an overall accuracy of 98.44%, the segmentation network attained a mean pixel accuracy of 90.44%, and the derived cloud cover showed excellent agreement with the manually annotated reference masks with a correlation coefficient of 0.92. Ablation studies further validate the effectiveness of the proposed deep-learning framework in improving the segmentation of complex cloud structures. The truncated EfficientNet-B0 encoder yields the most pronounced contribution, while the ASPP module improves the delineation of complex sky and cloud structures. When applied to the observations from the ground-based TIR all-sky camera, the retrieved cloud cover agreed well with the FY4B/AGRI cloud mask product, with a correlation coefficient ranging from 0.81 to 0.89 across the four seasonal months in 2023. These results prove that the proposed framework provides an effective and robust solution for long-term ground-based cloud monitoring using wide-field TIR whole-sky imagery. It also offers a practical foundation for applications such as cloud climatology, solar radiation assessment, and atmospheric observation.